Wind Power Forecasting

The science & practice

RAMESH M P
HOD, Wind Division, NAL
FORMER EXECUTIVE DIRECTOR, NIWE
Currently associated with Vaayu Power Corporation
Presentation plan
• The need
• Science behind
• The path this far
• CERC & SERC regulations
• What needs attention
• Way forward
Background to wind power forecasting
• Effective and efficient grid management requires a fairly good idea
about possible power flows into the grid and their consumption on
very short time frames of minutes to weeks and months for planning.
• Forecasting thermal or hydro power is simpler in short horizons.
• When grid penetration was low it did not matter much as systems
were large enough and the grid frequency variations were forgiven.
• As the RE penetration in terms of installed capacities started
increasing and at the same time grid frequency bands started getting
narrower, very quickly the regulations were brought in with ever
increasing demands on predictability.
The need
• Currently India has considerable share of non major hydro RE power on grid that is
weather dependent. It is going to grow further.
• The Indian Grid has to be able to deal with higher variability, storage technologies,
reprograming hydro power & development of decision making algorithms.
• Weather conditions have substantial variability.
• Weather modeling is a complex science and is not able to provide closed form solutions
presently as is being demanded.
• It is important to understand the nuances of the state of art and work towards best
possible way of dealing with uncertainties.
• It is not enough to be able to get estimates with small uncertainties for future time steps.
• It is important to have ability to absorb as much RE power as is feasible economically
instead of putting entire onus on RE generator.
The Science behind forecasting
• There are three critical steps
• Global Weather modeling
• Regional forecasting
• Down scaling & power modelling

• Global Weather modeling – As the name suggests – is carried out for entire earth
at grid points that could be with several kilometer spacing once in six hours.
• Input data to the model are taken from met services around the world.
• Notwithstanding guidelines the input data can be very sparse both temporally and spatially
• Quite a few assumptions, assigned values are used in such cases.
• The forecasts in such areas would have large uncertainties.
• Global models are known to work better in upper latitudes.
• Because of these factors results from the global models can not be taken as
sacrosanct in our context.
Forecasting State of Atmosphere in future steps
• Meteorological observations are carried out with a view to be able to
predict development of atmospheric conditions over short time scales
(minutes to weeks).
• These observations are made at pre-fixed times by meteorologists around
the world.
• The data thus collected would be used to calculate what will be the state of
atmosphere in near future using mathematical models.
• Essentially atmospheric modeling is rooted in equations of mass and
momentum conservation (Navier Stokes equations).
• Before we get into this, some background understanding will be helpful.
CHAOS – reality of atmospheric motions
• Chaotic systems are hyper sensitive to inputs given to
them.
• Any measurement is always associated with an
uncertainty. This point should be kept in mind
• Chaos being what it is, it has to be understood that
Accuracy of Input data is of great importance.
• Discovery of Chaos in atmosphere is attributed to Lorenz
who stated:
• “The Chaotic nature of atmospheric flow has been
popularly described as when a butterfly flaps its wings in
Brazil, it could lead to a tornado in Texas”. In other words
the outcome of model outputs can be widely divergent
for imperceptible changes in the input variables.
• While it may not be as dramatic, it essentially implies is
that there is a lot of unpredictability & uncertainty in the
way atmospheric flow develops in time and space.
Global modelling of atmospheric motions
Regional forecasting
• This kind of forecasting is a sub set of global modelling and can be down
scaled to few hundred meter spaced grids. This may provide a little bit
better information but could continue to carry the issues forward in results.
• However, based on radar based observations Satellite based inputs and
other inputs some improvements – in terms extreme events such as
thunderstorms can be captured in advance. However for precise location
specific quantitative forecasting is still some distance away.
• There is a possibility of running High Resolution Rapid Refresh (HRRR)
models that can be run every hour and making available outputs at one
hour or less interval. This is yet to happen for use with RE in India
 DOWN SCALING TO WIND FARM SITE

Input data from surface station & other sources

Wind farm

NWP Output at grid points

•Inputs to NWP from Observatories (white dots)

•NWP outputs at Grid points as Synthesised data sets (red dots)
•Nested (Sub-nets) calculations over project area based on local terrain, model training etc. (White patch)
•This data used in conjunction with on site SCADA to Produce future power sequences.
Wind Power Forecasting system

Probabilistic forecasts
Normally this is hourly elsewhere
Deterministic & Probabilistic forecasting
• When the inputs to the forecast model are
well defined and complete with no
uncertainty, the model is expected to
output with very low uncertainty. Such
forecasts are known as deterministic
forecasts.
• In real world this does not happen. In
order to deal with this ensemble
calculations are carried out. The input
variables are varied ever so slightly and a
large number of forecasts are obtained.
• There can be some convergence of the
output which can be used along with
probability quantiles defined
Ensemble Prediction
• The forecast model is run
hundreds of times and results so
obtained are to be dealt with
using probability
• There will be a likelihood that
there will be a convergence of
the results at future time steps.
This would define possible state
of atmospheric flow at future
time steps.
 Critical Ramp Events
• Ramp events are most often quoted difficulties in absorbing
RE power.
• It actually depends on the entire system’s ability to take them
in their stride at the point of common connection.
• Typically ramping happens over 2 to 4 hours at a ramp rate of
10/15% over 2 to 6 hours.
• If we look at the impact of these swings from the total power
flow on the grid at that point, it may not be seen as a major
grid event. But it is never reported in that manner.
• Forecaster must be able to take these into account as special
cases
• Aggregation will largely even out such situations.
• These events need to be treated using extreme event
analysis and require probabilistic methods with time and
space dependency.
 Indian Wind Power
Forecasting & Scheduling
For India
system

Info from
Satellite etc.
Ensembles, General
Spatial grid based
Generator has no computations
role to play here

Generator’s
role begins & QCA
ends here
Process of creating forecasts & schedules in India
Forcaster
•NWP
•Static data

Generator Output to SLDC

•SS metering data
•SCADA
•AvC

Inputs:
•TRAINED MODELS

• Day Ahead
• SCADA
• Revisions
• SS metering data
• RECONCILIATIONS
• AvC QCA
INTERACTS WITH
GENERATORS,
FORECASTERS & SLDC
Practical issues while forecasting
• NWP outputs typically come every six hours at predefined grid points.
• The forecast revisions can happen every sixth time block and there is
little that the forecaster can do to capture serious ramps to update.
• The other issue is how to deal with grid security based backing down
of the wind farm.
• In Day ahead forecast which needs to be provided by 10/11 AM of the
previous day, for data starting at 00:00 hours to 23:45 hours, the
forecast is already 14 hours old. A certain amount of divergence in
results would have already occurred as time advances from zero point
Typical issues of forecast v/s actual trajectory
• The forecasts, due to a variety of
Level error
reasons come with two significant Actual
errors:
• Phase error – where the time stamp
of given forecast can be either
advanced or retarded. In both cases
the calculated errors would be Phase
significant notwithstanding the fact Predicted
error
that predictions are close.
• Level error – could be a case of over
prediction or under prediction.
• Both can happen together.
CERC/SERC regulations
• The regulators started looking at forecasting needs as early as 2009.
By 2014 regulations were announced.
• Since then, a number of revisions have been made based on an
assumption that since it has been going on for so long, things must
have improved considerably.
• In reality there has been no perceptible improvement in the forecasts
produced nor is there adequate consistency as the basic issues have
not been addressed at all.
DSM charges
Penal charges as a percentage of net income
• Based on CERC regulations from DRAFT CERC
2014 to draft regulations of 2024 Substation Capacity 2014 2022 2023 2024

there has been an increase in Akal/Jajiya

Akal Bhu
230 MW
453 MW
0.48%
0.41%
4.35%
4.53%
1.60%
1.52%
2.67%
2.34%
DSM charges. Andhra lake
MH 106 MW 0.49% 4.32% 1.56% 2.67%
• The regulations need to take into Sadodar 462 MW 0.19% 3.28% 0.69% 1.22%

the state of art of forecasting

science.
• The basis of creating forecasts
and schedules is the weather
prediction science which works
with limited measured data.
It can be seen that the DSM charges in
different situations has been increasing with
each revision of regulations. Basically it
shows a disconnect with the advances in
forecasting sciences.
Error quantification & forecast value
• There are quite a few firms dealing with forecasting.
• Most of them promise to provide forecasts based on very advanced, AI and
ML enabled solutions.
• Unfortunately benchmarking these services is very complex.
• With attempts to minimize DSM charges good forecasting takes a back
seat.
• Attention needs to be paid towards bringing in some standardization into
the field.
• There are IEA recommendations which could serve as a starting point
• Some of the key indicators of the how well the models are performing
needs to be looked at.
How do we deal with Forecast errors
• With the current methods of error evaluation using the IEA
recommendations becomes difficult.
• As of the moment even if chosen service provider is consistently
giving bad forecasts, the rules are so complex that the generators
hesitate to change them.
• It will be necessary to deal with these aspects more holistically and
firmly.
• The following slides provide some of the basic definitions that would
help in evaluation.
International practices have defined some ways of doing this. At this time we have kept the
process simple as per CERC guidelines and not really thinking beyond them.

Forecast error is defined by

𝑒𝑒𝑡𝑡+𝑘𝑘|𝑡𝑡 = 𝑃𝑃�𝑡𝑡+𝑘𝑘|𝑡𝑡 − 𝑃𝑃𝑡𝑡+𝑘𝑘

Where,

et+k|t = is the error corresponding to time t+k for prediction made at time t, where k is the forecast
horizon

Pt+k is the power measured at time t+k

𝑃𝑃�t+k|t is the predicted power for time t+k

CERC regulation has defined the percentage error as

𝑒𝑒 𝑡𝑡+𝑘𝑘|𝑡𝑡
. 100 %, AVP= Available installed power
𝐴𝐴𝐴𝐴𝐴𝐴

Source:- Prediction Error-Based Power Forecasting of Wind Energy System Using Hybrid WT–ROPSO–NARMAX
Model Aamer A. Shah et al
23
Normalized error is defined by:
𝑁𝑁𝑒𝑒𝑡𝑡+𝑘𝑘|𝑡𝑡
Where Pinst is the installed power
CERC has taken cognizance
𝑒𝑒𝑡𝑡+𝑘𝑘|𝑡𝑡 𝑃𝑃𝑡𝑡+𝑘𝑘 − 𝑃𝑃�𝑡𝑡+𝑘𝑘|𝑡𝑡 of the fact that the
= = operational installed
𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 capacity can be different
from the Installed capacity
due to a variety of reasons.

Mean Average error for a forecast horizon k is given by:

𝑁𝑁
1
𝑀𝑀𝑀𝑀𝑀𝑀𝐾𝐾 = ��𝑒𝑒𝑡𝑡+𝑘𝑘|𝑡𝑡 �
𝑁𝑁
𝑡𝑡=1

where |et+k|t } is the absolute forecast error (both positive and negative errors)
defined above.
To make it feasible to compare different wind farms and forecast
methods, it is a practice to divide the MAE by installed capacity.
𝑀𝑀𝑀𝑀𝑀𝑀𝑘𝑘
𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑘𝑘 =
𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖
24
 Root Mean Square Error is defined by:

𝑁𝑁
1
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐾𝐾 = � � 𝑒𝑒𝑡𝑡+𝑘𝑘|𝑡𝑡
2
𝑁𝑁
1

Normalized Root Mean Square Error is given by:

𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝐾𝐾
𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝐾𝐾 =
𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖

Internationally these definitions are used to define forecast errors and are used
to compare outputs from different service providers

25
A re-look at NWP inputs and possible improvements
• One of the primary inputs that the models require is as much high
quality data as is feasible.
• The inputs to global circulation modelling comes from about 500
surface stations, balloon flights from less than 50 stations, satellite
based observations few other sources.
• There is a claim about private met stations for specific & special
purposes and that does not enter the common pool of data.
• While this may be adequate for some qualitative understanding of
state of the atmospheric flows, there is room for much improvement
Way forward
• With the background of forecast accuracies having tabled off it has
become essential to look at improvements at the root of the issues
we are facing.
• Two important sources of data have been proposed.
• Wind farm based SCADA can provide Wind Speed and Directions
• Installation of Automatic weather Stations may be considered.
• Chairperson, CEA set up a working group to make recommendations
for next course of action.
• The group headed by IMD supported by NCMRWF and other
Governmental establishments (MOP,MNRE etc.) and Associations
have made some progress.
Progress This far
• Since the task force was set up in February 2023 the committee has
met six times via Microsoft TEAMS and physical meetings.
• Sample wind farm SCADA data is being uploaded to IMD and
NCMRWF ftps with a granularity of 15 minutes.
• The data reaches their ftp with about 3 to 4 minutes latency and
uptime is >90%.
• Scientists have been looking at the data.
• Whatever clarifications needed have been provided.

28
Progress This far
• Discussions and follow up actions have resulted in
• Agreeing that this data, if available from all wind farm sites can make a
difference in improvement of NWP results.
• The need for other meteorological data continues to be an issue if we wish to
employ them in data assimilation.
• Some meteorological data is being collected at POSOCO substations and these
data must also be ported to NCMRWF
• For this to happen, Automatic Weather Stations also need to be installed at
strategic locations, continuously monitored and maintained.
• The data from AWS also must reach the modelers in time so that it could be
used at the assimilation level to obtain better results.
• It was also proposed that we should run a pilot study in Gujarat and
Rajasthan & subsequently expand to the entire country

29
Progress this far
• We have tried to put together the wind farms and corresponding
Substations.
• Initial estimates of number of wind farm specific substations is about 3200
substations.
• It is assumed that one substation of 33 kV/66/132/220/400 serves wind
turbines with in a radius of 25 to 30 km.
• The meteorological parameters captured at the respective sub stations are
expected to be representative for the wind farm area.
• It should be understood that this is much better than having no data at all
• The wind speed and direction information captured by Wind farm SCADA
will provide good scaling and possibly used in assimilation.

30
 Plenty of high
Windfarms across India quality onsite
data can become
available
We have windfarms at remote areas and the farm data
(including some meteorological data is available in near
real time.
This data is going abroad to the so called forecast service
providers in bits and pieces.
A dialogue with IMD was started a couple of years back
and some sample data is being shared with them for
study in near real time.
They have evinced serious interest in this as a possible
data source at INPUT level. This approach can be a game
changer.
It will need full support from the respective ministries.
It is to be hoped that the major departure from current
carrot & stick policy, with no carrot of course.

31
Substation distribution

32
Wind turbines &
Substations super-
imposed

33
Pilot study in Gujarat

25 km radius circles
showing possible
distribution of AWS
For Gujarat we envisage
about 23 AWS deployment
at the Substations
A similar exercise can be
carried out for other
states.
There is a possibility of
porting met data collected
at PGCIL and other
substations if they are
available.
34
Some considerations
• In plain terrain the meteorological parameters can have a spatial
distribution of about 25 km radius.
• For Gujarat we would need about 23 AWS’es.
• All over India we may think of about 100/150 AWS’es for this purpose.
• We need to see how best to deal with this aspect.
• These AWS need to be GPRS enabled.
• They would need security & some routine O & M.
• There should be a system to manage them properly.
• Typically they would cost 3 to 4 lacs/station and perhaps some central
assistance will help.

35
Structuring the project
• There has to be a structured system in place that will take care of
roles and responsibilities to see that there is no break in service.
• IMD/NCMRWF may decide the instrumentation needed, finalize
requirements. It will be best if they provide the systems under some
formal arrangement. Else it should be fully specified and validated. If
MNRE/NIWE provides administrative support, it may be ideal.
• Perhaps Wind Generators will take care of installation, commissioning
of the instruments. Take care to see that data flow shall not stop.
• Some Governmental Directions will be necessary to push the project
further.
36
Recommendations
• Getting the Wind farm SCADA based wind speed and direction
information to NCMRWF.
• Finalizing economic but effective AWS specs and modalities of
deployment.
• Testing the concept in Gujarat & Rajasthan.
• Creating a structure for the complete activity with roles and
responsibilities for sustained operation.
• We will require suitable directions to all stake holders to make this
happen.

37
Conclusions
• There can be vast improvements in the NWP outputs if the field data
from SCADA & on Site meteorological data are used as inputs – after
due verification & validation.
• Improved NWP results would automatically improve the forecast
accuracy.
• We must consider setting up High Resolution Rapid Refresh
forecasting systems that would help the forecasts get refreshed every
hour instead of current practice of 6 hour forecasts with low
resolution
• The forecasts must be carried out with state level aggregation instead
of merely collecting individual forecasts.
Conclusions
• The Indian Grid has to be able to deal with higher variability, storage
technologies, reprograming hydro power & development of decision
making algorithms. This initiative is a step in that direction.
• This public –private partnership is perhaps one of the best ways of
ensuring higher grid integration of wind power without compromising
safety.
• It is not a one time effort. It is something that has to be built and
sustained in the long run.
• This program has the potential to become a Make in Bharat if all stake
holders work towards making this happen.
40